Training VGGT from scratch

[stephanie-fu]

Hi, thank you for sharing such great work! I'm interested in the from-scratch training recipe for VGGT, and had a few clarification questions on the config:

• the paper mentions that 64 A100s were used, with [2,24] frames sampled per sequence and "maintaining the constant total of 48 frames within each batch"
  • does this mean that the effective batch size was 64 gpus * (48 frames/gpu) / ([2, 24] frames/sequence) = [128, 1536] sequences? Or were the samples reduced in a different way?
• is there a general range of loss magnitudes that we should expect to indicate proper convergence at the end? Especially for the tracking loss term, as there is not a clean version integrated into the codebase yet.
• was data from the list here Complete list of training datasets #36 sampled uniformly over the course of training?

[jytime]

Hi @stephanie-fu ,

Thanks for the interest.

1. Yes that's correct.
2. You can clearly see that the curve tends to be flat when the model converges.
3. You can start from a uniform sampling, that's okay. Based on our explorations later, we suggest controlling a ratio of 4:6 for real and synthetic datasets